Evaluating the impacts of interventions on different features of an epidemic curve through random forest metamodels

Simulation modeling is widely used to investigate the impact of interventions on infectious disease transmission. However, individual-based models often involve numerous parameters, making exhaustive exploration computationally prohibitive. Metamodels can address this limitation, and when coupled with interpretable machine learning, they can provide insights into intervention impacts. Moreover, multiple features of epidemic curves should be considered when evaluating interventions. Using the open-source agent-based model Covasim, we examined non-pharmaceutical interventions affecting key epidemic curve features, such as peak day, peak value, and attack rate. We employed random forests as a metamodeling technique to capture nonlinear input–output relationships. Then, we used the Boruta algorithm to identify impactful interventions. Results showed that random forest metamodels explained moderate to high variance across outputs. The Boruta method indicated that influential interventions differed by outcome. Social distancing measures, such as school closures and workplace density reductions, were mostly unimportant in the current Covasim setting, whereas intervention timing, testing, contact tracing, and mask use consistently influenced all outputs to varying degrees. Our metamodel- and variable selection-based framework can provide insights into the effectiveness of multiple interventions across various characteristics of the epidemic curve, thereby supporting public health decision-making at the onset of an epidemic.